Multi-modal data fusion spot electricity price prediction method

Abstract

The invention discloses a spot electricity price prediction method based on multi-modal data fusion, particularly relates to the field of electricity price prediction, and is used for solving the problems of error accumulation amplification and prediction precision distortion caused by high-frequency noise in multi-modal data fusion. Direct interference of noise on subsequent processes is avoided; extracting and dynamically revising confidence information to ensure that the reliability of different components is accurately evaluated; secondly, a high-confidence part is processed preferentially in the fusion stage, and a low-confidence part is smoothed, so that a core trend is reserved, and fluctuation deviation is reduced at the same time; and finally, iteratively correcting the deviation by using the load reference until the convergence and outputting a prediction curve. The front-end division provides pure input, the middle-section revision and fusion interaction optimization weight, and the rear-end correction refines the result through feedback circulation, so that high-frequency noise is effectively suppressed in new energy electricity price prediction, market dynamic changes are captured, and the accuracy and stability of a prediction curve are improved.

Description

Technical Field

[0001] This invention relates to the field of electricity price forecasting, and more specifically, to a method for forecasting spot electricity prices using multimodal data fusion. Background Technology

[0002] In the paper "A Day-ahead Electricity Price Forecasting Method for the New Energy Power Market Based on Multi-Source Information Fusion" (application number CN117911076A), the spot electricity price is first decomposed into multiple intrinsic modal components and residual terms through complementary set empirical mode decomposition. Subsequently, each component is input into a sparrow search-least squares support vector machine model to generate corresponding prediction curves. All outputs are then simply added together at the terminal to directly form the final electricity price trend. This process highly emphasizes decomposition and single-modal modeling, but fails to distinguish the importance and error characteristics of different modes in the final fusion stage.

[0003] Multimodal components naturally differ in their sources, fluctuation amplitudes, and prediction difficulty: high-frequency components from wind and solar power output are noisy, while low-frequency components from load and matching signals are relatively stable. When all components are superimposed with equal weights at the terminal, the large bias carried by high-frequency noise is introduced into the overall curve without thresholding, leading to error accumulation and amplification. This phenomenon arises because the fusion layer lacks a dynamic adjustment mechanism for component confidence; in the process, high-frequency components experience drastic jumps within a short time interval, which the prediction model cannot fully capture, and their errors are simultaneously introduced into the result with the superposition step; in the result, the overall electricity price prediction accuracy is distorted synchronously with the noise peak, affecting the reference value for market clearing. This problem is not self-evident because existing research focuses on front-end feature weighting and decomposition algorithms, while paying insufficient attention to the transmission mechanism of back-end multimodal fusion errors; with the increasing proportion of new energy, if a differentiated fusion strategy is not established, prediction bias will continue to expand, directly weakening the practicality and credibility of multimodal data fusion methods. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a spot electricity price forecasting method based on multimodal data fusion. This method separates steady-state and disturbance signals by dividing the instantaneous bandwidth of the multimodal sequence, avoiding direct noise interference in subsequent processes. Next, confidence information is extracted and dynamically revised to ensure accurate assessment of the reliability of different components. Then, in the fusion stage, high-confidence components are processed first, while low-confidence components are smoothed to preserve the core trend while reducing fluctuation deviations. Finally, deviations are iteratively corrected using a load benchmark until the output forecast curve is converged. The front-end segmentation provides clean input, the mid-stage revision and fusion interactively optimize weights, and the back-end correction refines the results through feedback loops. This effectively suppresses high-frequency noise, captures dynamic market changes, and improves the accuracy and stability of the forecast curve in renewable energy electricity price forecasting, thereby solving the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] S1: Real-time extraction of multi-modal sequences such as wind and solar power output, load and matching signals, adaptive division into steady-state bundles and disturbance bundles based on instantaneous bandwidth and then writing them into the component pool;

[0007] S2: Extract confidence information from the short-window reconstruction error curve of each sequence in the component pool and write it into the confidence mapping table;

[0008] S3: Based on phase consistency characteristics and amplitude mutation characteristics, a revision factor is output through a multi-core support vector machine to refresh the confidence mapping table level in real time;

[0009] S4: First, overlay the high-confidence components according to the refreshed confidence mapping table, and then perform phase stretching smoothing on the low-confidence components before overlaying to generate the first round of fusion curve;

[0010] S5: The reference curve is output by the load-driven fast reference model. The difference between the curve and the first-round fusion curve is used to obtain the deviation vector. The confidence mapping table is refreshed with the deviation vector and the next round of step S4 is triggered. When the deviation vector converges to the predetermined limit, the prediction curve is output and the confidence evolution trajectory is written back for subsequent retrieval.

[0011] In a preferred embodiment, step S1 includes the following:

[0012] The power output sequence, load sequence, and matching signal sequence are updated every second through the data interface to form a continuous multimodal sequence set. The instantaneous bandwidth of each sequence in the multimodal sequence set is calculated using Hilbert transform to quantify the intensity of local fluctuations. An adaptive threshold is calculated based on the instantaneous bandwidth to determine the sequence segment assignment. Sequence segments with instantaneous bandwidth less than or equal to the threshold are assigned to the steady-state bundle, while those with bandwidth greater than the threshold are assigned to the disturbance bundle. The steady-state bundle and the disturbance bundle are written into the component pool.

[0013] In a preferred embodiment, step S2 includes the following:

[0014] Each sequence from the steady-state and perturbation bundles is read from the component pool to form a set of sequences to be processed. The amplitude change rate of the local interval of the sequence is calculated to dynamically set the window length to adapt to the fluctuation scale. Hilbert transform is applied to the sequence sample points within the window to generate analytical signals and extract instantaneous amplitude and phase. The instantaneous amplitude is mapped according to the instantaneous frequency distribution to generate a local Hilbert spectrum. The phase integral and amplitude product of the selected principal mode component are summed through inverse transform to reconstruct the approximate sequence. The root mean square error is used to calculate the reconstruction error curve to quantify the deviation between sequences.

[0015] In a preferred embodiment, step S2 further includes the following:

[0016] The confidence score is extracted by taking the reciprocal of the error curve through integral reconstruction, and the sequence window corresponding to the confidence score is written into the confidence mapping table.

[0017] In a preferred embodiment, step S3 includes the following:

[0018] The original sequence corresponding to the confidence score array of each sequence is read from the confidence mapping table to form a feature extraction sequence set; the inverse sequence of the generated sequence is normalized by cross-correlation operation and the maximum value is taken to extract the phase overlap to quantify the phase matching degree; the ratio of the difference between the mean of the amplitude sequence and the high quantile value is calculated in the same window to extract the amplitude jump rate to quantify the amplitude deviation.

[0019] In a preferred embodiment, step S3 further includes the following:

[0020] Using phase overlap and velocity jump rate as input vectors, a multi-kernel support vector machine is trained to minimize risk, solve for support vectors, and calculate the inner product, outputting the dynamics of the revision factor response features; the revision factor is multiplied by the confidence score array to refresh the confidence mapping table.

[0021] In a preferred embodiment, step S4 includes the following:

[0022] Read the confidence score array of all sequences from the refreshed confidence mapping table to form a component classification set. Sequences with confidence scores higher than a preset threshold are classified as high-confidence components, and sequences with confidence scores lower than or equal to the preset threshold are classified as low-confidence components.

[0023] In a preferred embodiment, step S4 further includes the following:

[0024] The high-confidence components are added point by point to generate a preliminary curve to accumulate a reliable trend. The phase sequence of the low-confidence components is extracted, and the inverse of the local curvature integral is calculated to obtain a phase scaling factor multiplied by the phase cosine to reconstruct a smooth sequence to compress the oscillation amplitude. The smoothed low-confidence components are added to the preliminary curve to generate the first-round fusion curve to integrate the component contributions.

[0025] In a preferred embodiment, step S5 includes the following:

[0026] The load-driven fast reference model is activated, using multinomial regression to capture load trends and output a reference curve. This curve is then subtracted point-by-point from the initial fusion curve to generate a deviation vector, which quantifies the error distribution. The confidence map is refreshed using a locally integral negative exponential decay confidence score array based on the deviation vector. This refresh triggers the next step, S4, where the updated confidence map is used to re-execute the fusion process.

[0027] In a preferred embodiment, step S5 further includes the following:

[0028] Calculate the overall magnitude of the deviation vector, compare it with the predetermined limit set by the median value of the historical magnitude integral, if the overall magnitude is greater than the predetermined limit, continue iterating, if it is less than or equal to the predetermined limit, output the current fusion curve as the prediction curve, and write back the confidence evolution trajectory.

[0029] The technical effects and advantages of the multimodal data fusion spot electricity price forecasting method of this invention are as follows:

[0030] This invention separates steady-state and disturbance signals by dividing the instantaneous bandwidth of a multimodal sequence, avoiding direct noise interference in subsequent processes. Next, it extracts and dynamically revises confidence information to ensure accurate assessment of the reliability of different components. Then, in the fusion stage, it prioritizes the processing of high-confidence components and smooths low-confidence components to preserve the core trend while reducing fluctuation deviations. Finally, it iteratively corrects deviations using a load benchmark until the output prediction curve is converged. These steps collaborate: the front-end partitioning provides clean input, the mid-stage revision and fusion interactively optimize weights, and the back-end correction refines the results through feedback loops. This effectively suppresses high-frequency noise, captures dynamic market changes, and improves the accuracy and stability of the prediction curve in renewable energy price forecasting, providing more reliable support for power trading decisions. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the spot electricity price prediction method based on multimodal data fusion of the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Example 1: Figure 1 This invention presents a spot electricity price forecasting method based on multimodal data fusion, comprising:

[0034] S1: Extract multi-modal sequences such as wind and solar power output, load and matching signals in real time, and adaptively divide them into steady-state bundles and disturbance bundles according to instantaneous bandwidth before writing them into the component pool.

[0035] S2: Extract confidence information from the error curve of each sequence in the component pool using a short window and write it into the confidence mapping table.

[0036] S3: Based on phase consistency characteristics and amplitude mutation characteristics, a revision factor is output through a multi-kernel support vector machine to refresh the confidence mapping table level in real time.

[0037] S4: First, overlay the high-confidence components according to the refreshed confidence mapping table, then perform phase stretching and smoothing on the low-confidence components before overlaying to generate the first round of fusion curve.

[0038] S5: The reference curve is output by the load-driven fast reference model. The difference between the curve and the first-round fusion curve is used to obtain the deviation vector. The confidence mapping table is refreshed with the deviation vector and the next round of step S4 is triggered. When the deviation vector converges to the predetermined limit, the prediction curve is output and the confidence evolution trajectory is written back for subsequent retrieval.

[0039] In the field of spot electricity price forecasting in the new energy power market, spot electricity prices are influenced by multimodal data such as wind and solar power output, load, and matching signals. These data come from diverse sources, resulting in price sequences containing both steady-state trends and disturbances. If these are not distinguished and fused, high-frequency disturbances can amplify prediction errors and reduce the reliability of market decisions. To address the problem of neglecting modal differences in the fusion stage of existing methods, this invention introduces a multimodal data fusion strategy. First, step S1 extracts and adaptively segments the multimodal sequence in real time to separate steady-state and disturbance components, preventing noise from directly propagating to subsequent predictions. This step serves as the foundation for end-to-end front-end noise reduction, ensuring that subsequent confidence assessments and fusion are optimized for different bundle characteristics, thereby improving overall prediction accuracy and robustness.

[0040] 1-1. Real-time acquisition of wind and solar power output sequences, load sequences, and matching signal sequences. These sequences represent the values ​​of wind and solar power output, load, and matching signal at each time point, respectively. These sequences are updated every second through the data interface to form a continuous multimodal sequence set. This allows subsequent segmentation based on instantaneous bandwidth to directly reference the content of the multimodal sequence set, thereby capturing the real-time nature and diversity of data sources during the extraction stage. This ensures that the segmentation process reflects the actual market dynamics and avoids distortion caused by static processing.

[0041] This invention selects only wind and solar power output sequences, load sequences, and matching signal sequences in its spot electricity price forecasting method because these data represent the core driving factors of the renewable energy power market: the wind and solar power output sequence captures the instantaneous fluctuation characteristics of the renewable energy supply side, the load sequence reflects the dynamic changes on the demand side, and the matching signal sequence reflects the supply and demand balance and price formation process under the market trading mechanism. By focusing on these three multimodal data, the main information sources required for electricity price forecasting can be efficiently covered, avoiding the introduction of redundant variables that would increase model complexity or cause noise interference. At the same time, it ensures that the forecasting process accurately integrates high-frequency disturbances and low-frequency trends in the context of the increasing proportion of renewable energy, thereby improving the practicality and accuracy of the overall method.

[0042] 1) The acquisition and processing of wind and solar power output sequences involves real-time monitoring of parameters such as wind speed, wind direction, solar radiation intensity, and power output of generating equipment through sensor networks installed in wind farms and photovoltaic power stations. These parameters are converted into digital signals by analog-to-digital converters and then aggregated every second by a dedicated data acquisition unit to form continuous time series data. Subsequently, noise filtering is performed on the acquired raw data, a low-pass filter is used to remove high-frequency interference, and an interpolation algorithm is applied to fill in missing values ​​to ensure sequence integrity. Finally, the processed sequence is standardized to a unified dimension to facilitate fusion with other modal data, thereby accurately reflecting the instantaneous fluctuation characteristics of new energy power generation.

[0043] 2) The load series acquisition and processing process utilizes smart meters and substation monitoring equipment deployed at the power grid user end to record information such as total electricity load, peak-valley distribution, and regional load changes every second. This information is transmitted to the central data center via wireless or wired communication protocols to form continuous time series data. Next, anomaly detection algorithms are applied to the raw data to remove outliers caused by equipment failures or sudden events, and moving average smoothing techniques are used to reduce the impact of short-term noise. Finally, the series is time-series aligned and normalized to match the sampling rate of other series, ensuring that market demand dynamics are reliably reflected in spot electricity price forecasts.

[0044] 3) The acquisition and processing of the matching signal sequence relies on the real-time trading system of the power trading platform. Every second, it captures trading signals such as bidding quotations, transaction volume, and matching results during the market clearing process. These signals are transmitted to the input interface of the prediction model through an encrypted channel, forming continuous time series data. Then, the acquired data is verified and checked to confirm that there is no tampering or delay. Differential calculation is used to highlight the trend of signal changes and filter out invalid matching events. Finally, the sequence is converted into a standardized format to facilitate integration with other multimodal data, thereby accurately capturing the instantaneous adjustment mechanism of market supply and demand balance.

[0045] 1-2. Calculate the instantaneous bandwidth of the sequence in each multimodal sequence set using the Hilbert transform. The Hilbert transform is a mathematical operation that converts the original sequence into an analytic signal and can extract the instantaneous characteristics of the sequence, including amplitude and phase information.

[0046] First, apply the Hilbert transform to each sequence in the multimodal sequence set to obtain a composite signal containing the original sequence and the result of the imaginary part transformation, i.e., an analytic signal, where the imaginary part is an orthogonal component obtained by integrating the original sequence.

[0047] Then, the instantaneous frequency is obtained from the analytic signal by calculating the derivative of the phase of the analytic signal and dividing it by twice pi, which represents the oscillation rate of the sequence at a local time point. Then, the instantaneous bandwidth is calculated by taking the derivative of the natural logarithm of the magnitude of the analytic signal, squaring it, adding the square of the derivative of the instantaneous frequency, and then taking the square root, thereby quantifying the local fluctuation intensity of the sequence at each time point.

[0048] This calculation method emphasizes the joint evaluation of amplitude and frequency changes, and can sensitively capture the difference between high-frequency disturbances and low-frequency trends. Therefore, it is naturally suitable for distinguishing between steady state and disturbance in the processing of multimodal sequence sets, and ensures that the instantaneous bandwidth can dynamically respond to sudden variations caused by new energy output when used as the basis for division.

[0049] 1-3. The threshold is adaptively set based on the calculated instantaneous bandwidth. The principle is to add the average value of the instantaneous bandwidth at all time points to the maximum instantaneous bandwidth and multiply it by an adjustment coefficient. This adjustment coefficient is used to balance the sensitivity to extreme fluctuations and the overall stability, and to avoid the division deviation caused by the threshold being too high or too low. If the instantaneous bandwidth at a certain time point is less than or equal to the threshold, the corresponding sequence segment is assigned to the steady-state bundle; otherwise, it is assigned to the disturbance bundle.

[0050] This adaptive thresholding mechanism integrates average and extreme value information to achieve dynamic classification of multimodal sequence sets. This allows the steady-state bundle to include low-volatility components to preserve the basic market trend, while the perturbation bundle includes high-volatility components to isolate noise. This improves the targeting of the fusion stage after segmentation and reduces error amplification.

[0051] Finally, the divided steady-state bundles and disturbance bundles are written into the component pool, which stores the steady-state bundles and disturbance bundles in a structured form. This facilitates subsequent steps to directly read each sequence from the component pool for short-window reconstruction error curve extraction. This writing operation ensures the complete link of data from extraction to division, forming traceable component pool content, and providing a differentiated input basis for spot electricity price forecasting.

[0052] Step S1 has divided the multimodal sequence into steady-state bundles and perturbation bundles and written them into the component pool, thereby separating the low-frequency trend and high-frequency noise, laying a differentiated foundation for subsequent processing. However, in order to further quantify the reliability of the sequence in each bundle to avoid noise dominance during fusion, step S2 needs to extract confidence information by reconstructing the error curve through a short window, ensuring that the confidence mapping table can dynamically reflect the error characteristics of the components, thereby achieving targeted adjustment in multimodal fusion.

[0053] 2-1. Read each sequence from the steady-state bundle and the perturbation bundle from the component pool to form a set of sequences to be processed. The steady-state bundle sequences reflect the reliability of low-frequency trends, while the perturbation bundle sequences highlight high-frequency uncertainties.

[0054] 2-2. Apply the short window reconstruction method to each sequence in the set of sequences to be processed. The short window reconstruction method is a reconstruction technique based on local signal decomposition. It can simulate the intrinsic mode of the sequence within a limited window. First, determine the window length. Its adjustment principle is to dynamically set it by calculating the amplitude change rate of the sequence in the local interval. The amplitude change rate is obtained by summing the absolute values ​​of the differences between adjacent points of the sequence and dividing by the number of points. This allows the window to expand in the steady-state bundle to cover the stable segment and shorten in the perturbation bundle to isolate jumps, thereby adapting to the fluctuation scale of different bundles.

[0055] 2-3. Perform Hilbert transform on the sequence sample points within the window. The principle of this transformation is to combine the original sequence with the imaginary part sequence obtained through integral orthogonal transformation to form an analytic signal. The analytic signal is used to extract the instantaneous amplitude as the envelope of the sequence and the instantaneous phase as the phase function of the sequence.

[0056] Then, a local Hilbert spectrum is generated, which is formed by mapping the instantaneous amplitude to the instantaneous frequency distribution, representing a three-dimensional representation of frequency-time-amplitude within the window;

[0057] The approximate sequence is then reconstructed through inverse transformation. The principle of inverse transformation is to sum the phase integral and amplitude product of the selected principal mode components in the Hilbert spectrum to recover a sequence form close to the original. This reconstruction emphasizes the preservation of the dominant frequency and can effectively simulate the trend of the steady-state bundle while exposing the residuals of the perturbation bundle.

[0058] Finally, the reconstruction error curve is calculated. The principle of the operation is to square the difference between the original sequence and the reconstructed sequence at each point, sum them up, and then take the square root of the summation result for the entire window, so as to obtain a curve value that quantifies the overall deviation, ensuring that the error curve reflects the incompleteness of the reconstruction, especially the parts in the perturbation bundle that are difficult to simulate.

[0059] 2-4. Confidence information is extracted based on the reconstruction error curve. The operation principle is to integrate the reconstruction error curve from the starting point to the ending point within each window. The integration is approximated by accumulating the error curve values ​​point by point and multiplying them by the interval between points to obtain the total cumulative error. Then, the reciprocal of the total cumulative error is taken as the confidence score. This reciprocal transformation ensures that the score is inversely proportional to the error. Furthermore, the dimensionless processing of the integration maintains the consistency of the score. Thus, the low integral value of the steady-state bundle sequence corresponds to the high confidence score to emphasize its reliability, while the high integral value of the perturbation bundle sequence corresponds to the low confidence score to indicate its uncertainty risk.

[0060] 2-5. Write the window corresponding to each sequence in the sequence set to be processed based on the extracted confidence scores into the confidence mapping table. The confidence mapping table stores all information in a structure with the sequence identifier as the key and the confidence score array as the value. This facilitates subsequent steps to directly reference the confidence scores for revision driven by phase consistency features and amplitude mutation features, thus naturally transitioning to the dynamic refresh mechanism after confidence quantization.

[0061] Step S2 has extracted the confidence score by reconstructing the error curve through a short window and written it into the confidence mapping table, thereby quantifying the initial reliability of the steady-state bundle and the perturbation bundle sequences. However, in order to adapt to the dynamic changes of the sequence and avoid the fusion bias caused by static confidence, step S3 needs to output the revision factor through a multi-core support vector machine based on the phase consistency characteristics and amplitude mutation characteristics to realize the real-time refresh of the confidence mapping table, thereby ensuring that the confidence score reflects the phase stability and amplitude stability of the sequence.

[0062] 3-1. First, read the original sequence corresponding to the confidence score array of each sequence from the confidence mapping table to form a feature extraction sequence set. The steady-state bundle sequence has a higher confidence score to capture trend continuity, while the perturbation bundle sequence has a lower confidence score to indicate fluctuation uncertainty. This reading method ensures that the feature extraction closely follows the quantified content, thereby incorporating the error context into the phase consistency and amplitude mutation analysis, which is convenient for highlighting the difference between the phase stability and amplitude stability of the sequence.

[0063] 3-2. Extract phase consistency features, i.e. phase overlap, for each sequence in the feature extraction sequence set. Phase overlap is an index that measures the degree of overlap of phase information of sequences and can sensitively detect the phase continuity of steady-state bundles and the phase break of perturbation bundles.

[0064] First, the reverse sequence of the original sequence is generated by rearranging all the point values ​​of the original sequence from end to beginning. Then, cross-correlation is performed. The principle of cross-correlation is to multiply the corresponding point values ​​of the original sequence and the reverse sequence one by one at each possible lag position and sum them to generate a correlation value sequence. Next, normalization is performed by dividing each element of the correlation value sequence by the maximum autocorrelation value of the sequence itself at zero lag, ensuring that the normalized value range is limited to zero and one. Finally, the maximum value in the normalized correlation value sequence is taken as the phase overlap. The calculation idea is to quantify the phase matching degree between the original sequence and the reverse sequence by the maximum correlation peak, ensuring that the phase overlap is dimensionless and proportional to the phase consistency. Thus, the higher phase overlap of the steady-state bundle sequence emphasizes its phase reliability, while the lower phase overlap of the perturbation bundle sequence reveals its phase instability.

[0065] 3-3. Extract amplitude abrupt change features within the same window, namely the amplitude jump rate. The amplitude jump rate is a measure to evaluate the severity of changes in the amplitude of a sequence and can distinguish between amplitude jumps in a perturbed bundle and uniform amplitudes in a steady-state bundle.

[0066] First, the mean of the amplitude sequence is calculated by summing all the values ​​within the window and dividing by the number of points. Then, the high quantile is determined by selecting a specific value at the top of the amplitude sequence after sorting it. This selection emphasizes the focus on extreme amplitudes. Next, the ratio of the difference between the high quantile and the mean to the mean is calculated as the amplitude jump rate. The idea behind this calculation is to quantify the degree of amplitude deviation from the average by using a relative deviation ratio, ensuring that the amplitude jump rate is dimensionless and proportional to the intensity of the abrupt change. This allows the higher amplitude jump rate of the perturbation bundle sequence to highlight its amplitude risk, while the lower amplitude jump rate of the steady-state bundle sequence confirms its amplitude consistency.

[0067] 3-4. Using the extracted phase overlap and velocity jump rate as input vectors, a revision factor is output through a multi-kernel support vector machine (SVM). The multi-kernel SVM is a nonlinear regression model integrating a polynomial kernel and a radial basis kernel, capable of simultaneously capturing global trends and local details. First, the multi-kernel SVM is trained using the combination of phase overlap and velocity jump rate from historical sequences as input vectors, with the historical adjustment magnitude of the corresponding confidence score as the output label. Support vectors are solved by minimizing the sum of empirical risk and regularization terms. This solution process uses a quadratic programming optimization algorithm for iterative convergence. The support vectors are the subset of input vectors that contribute significantly to the decision boundary. The polynomial kernel operation... The principle is to add a constant to the inner product of two input vectors and then raise it to the power of the product to capture higher-order polynomial relationships. The radial basis function (RBF) kernel operation is based on the negative exponential function of the squared Euclidean distance between two input vectors, emphasizing local similarity. The multi-kernel operation achieves fusion by linearly combining the outputs of the polynomial kernel and the RBF kernel, ensuring the model's nonlinear mapping of phase and amplitude features. During runtime, the current phase overlap and velocity jump rate are input, and the model calculates the multi-kernel inner product of the input vector and support vectors plus a bias term, outputting a revision factor. The calculation idea is to avoid explicit high-dimensional calculations through kernel tricks, ensuring that the revision factor accurately responds to feature dynamics, thereby adjusting the confidence score to adapt to the phase and amplitude variations of the sequence.

[0068] Multi-kernel support vector machines are constructed by fusing multiple kernel functions in a linear combination. The fusion form is a weighted sum of multiple kernel functions, where the weights satisfy the condition that they are non-negative and their sum is one, in order to adapt to complex data distributions and improve classification or regression performance. The construction process first selects various kernel functions such as linear kernel, polynomial kernel, and Gaussian radial basis kernel, each kernel function corresponding to different feature modes or data sources. Next, the kernel matrix of each kernel function is calculated, which is the sample similarity matrix. Then, these kernel matrices are linearly combined according to weights to generate a fused kernel matrix, which is used as input for support vector machines.

[0069] Multi-kernel support vector machine (SVM) training optimizes the dual problem based on the fusion kernel matrix. The objective function is to maximize the product of the Lagrange multiplier and the label minus half of the product of the fusion kernel function among the multipliers, while subtracting an insensitive loss term. Constraints include the sum of the multipliers being zero and the multipliers being bounded between zero and the penalty parameter divided by the number of samples. The training steps include preparing the training set fusion kernel matrix and label vectors, and using a sequential minimum optimization algorithm to iteratively update the Lagrange multipliers until the Karush-Kuhn-Tucker condition is satisfied. In each iteration, all parameters except one pair of multipliers are fixed, and only the current pair of multipliers is optimized to find the extreme value. For example, in classification tasks, the pre-computed fusion kernel matrix is ​​used as input to the SVM model. After training, the support vector expansion is obtained, which is the prediction value of a new sample as the sum of the product of the support vector multipliers, the label, and the fusion kernel function plus a bias term, thus achieving classification or regression prediction for the test set.

[0070] Multi-kernel support vector machine (SVM) optimization focuses on jointly learning kernel weights and model parameters. It solves for support vectors by minimizing the sum of empirical risk and regularization terms, using grid search or golden section search to traverse the weight combination space. Simultaneously, it evaluates cross-validation performance metrics such as accuracy to select the optimal weights. The optimization process also includes adjusting the internal parameters of a single kernel, such as Gaussian kernel bandwidth or polynomial kernel order, and using duality gaps or weight change thresholds as stopping criteria to ensure that the iteration converges to the optimal solution of the convex optimization problem. For example, in regression tasks, multi-kernel weights are optimized to minimize prediction bias, while ridge regression terms are incorporated to avoid ill-conditioned kernel matrices and ensure the model's robustness to noisy data.

[0071] Multi-kernel support vector machine parameter settings include a penalty parameter to balance model complexity and training error; a larger value results in a tighter fit but is prone to overfitting. The insensitive loss bias defines the allowable error range and affects the number of support vectors. Kernel weights are initialized to a uniform distribution and then adjusted through optimization. Single-kernel parameters, such as Gaussian kernel bandwidth, control mapping sensitivity, while the polynomial kernel order determines the degree of nonlinearity. Iteration parameters include a maximum number of iterations of 500, a minimum number of iterations of zero, a numerical precision of 10 to the power of -8, a weight change threshold of 10 to the power of -2, and a duality gap threshold of 0.01. For example, in the classification of the ion sphere dataset, setting the penalty parameter to 100, the kernel function parameters to Gaussian kernel bandwidth sequences from 0.5 to 20, and the polynomial kernel order from one to three, and determining the optimal combination through cross-validation, improves prediction accuracy.

[0072] 3-5. The confidence map level is refreshed in real time using the output revision factor. The refresh principle is to multiply each score in the confidence score array by the revision factor of the corresponding sequence to generate an updated score array. Then, the updated score array is used to overwrite the original array in the confidence map table, so that subsequent steps can directly use the refreshed content to classify and combine the confidence components, thereby strengthening the adaptive control of the sequence dynamics in the revision mechanism.

[0073] Step S3 has driven the multi-core support vector machine to refresh the confidence map table through phase consistency features and amplitude mutation features, thereby adapting the confidence score to the sequence dynamics. However, in order to distinguish reliability and avoid noise dominance in the fusion stage, step S4 needs to prioritize superimposing high confidence components and apply phase scaling smoothing to low confidence components according to the refreshed confidence map table to generate the first round of fusion curve, thereby ensuring that the curve suppresses perturbation bias while preserving the trend.

[0074] 4-1. First, read the confidence score arrays of all sequences from the refreshed confidence mapping table to form a component classification set. Sequences with confidence scores higher than a preset threshold are classified as high-confidence components, and sequences with confidence scores lower than or equal to the preset threshold are classified as low-confidence components. The preset threshold is adaptively determined by calculating the median value of the integral of all confidence score arrays. The principle of the integral median value calculation is to accumulate each confidence score array point by point from the beginning to the end, multiply by the interval between points, sort them, and take the middle position value as the threshold to ensure that the classification is based on the overall distribution center. Thus, in the fusion, high-confidence components are prioritized to strengthen the trend dominance, while low-confidence components are controlled to reduce noise transmission.

[0075] 4-2. Direct superposition is performed on high-confidence components in the component classification set. This superposition is a point-by-point accumulation method that can integrate the contributions of reliable sequences. First, the values ​​of all high-confidence components are collected at each time point. Then, these values ​​are added together to generate a preliminary curve. The principle of operation is to achieve uniformity of value range by summing synchronously at time points, so as to ensure that the preliminary curve reflects the collective trend of high-confidence components. This allows the preliminary curve to establish a stable skeleton in the early stage of fusion to improve the curve's resistance to disturbances.

[0076] 4-3. Phase scaling smoothing is applied to low-confidence components in the component classification set. Phase scaling smoothing is a noise suppression method based on phase adjustment, which can compress the oscillation amplitude of low-confidence components. First, the phase sequence of the low-confidence components is extracted, and the phase function is generated using Hilbert transform. The principle of Hilbert transform is to combine the low-confidence components with the imaginary components obtained through orthogonal integral transformation to form an analytic signal. The arctangent of the analytic signal is then taken to obtain the phase sequence. Then, the phase scaling factor is calculated. The principle of phase scaling factor calculation is to integrate the local curvature of the phase sequence within a window and take the reciprocal. Curvature is represented by the absolute value of the first derivative of the phase sequence. The first derivative is approximated by dividing the phase difference between adjacent points by the time interval. The integral is achieved by accumulating point by point and multiplying by the interval, ensuring that the factor is inversely proportional to the curvature accumulation to quantify the phase curvature. Then, a scaling operation is applied to multiply the phase sequence point by point by the phase scaling factor to generate an adjusted phase. The original amplitude sequence is then multiplied by the cosine value of the adjusted phase to reconstruct the smoothed sequence. The calculation idea is to reduce the dense oscillation region by phase compression while maintaining amplitude information, so that the low confidence component contributes controlled fluctuations after smoothing to avoid curve distortion.

[0077] 4-4. The smoothed low-confidence component is superimposed with the preliminary curve to generate the first-round fusion curve. This superposition is a sequential additive integration that can merge the processing results of high and low components. First, the values ​​of the smoothed low-confidence component are collected at each time point. Then, these values ​​are added to the preliminary curve point by point. The operation principle is to achieve a complete curve through time-synchronized summation, ensuring that the first-round fusion curve covers the contributions of all components and has a consistent value range. This allows the first-round fusion curve to incorporate smoothed disturbances under a stable trend to optimize the accuracy of electricity price trends, facilitating subsequent steps to use the first-round fusion curve for deviation analysis.

[0078] Step S4 has generated the first-round fusion curve, thereby integrating the differential contributions of high and low confidence components. However, in order to correct potential biases and avoid noise residue, step S5 needs to iteratively refresh the reference curve output by the load-driven fast reference model to achieve the final output after bias convergence, thereby ensuring that the prediction curve achieves a fine balance between trend and disturbance.

[0079] 5-1. The first-round fusion curve has been formed by step S4. This curve initially integrates the confidence differences of multimodal sequences. To achieve iterative correction, a load benchmark mechanism is introduced. First, the load-driven fast reference model is activated to output a reference curve. The load-driven fast reference model is a simplified polynomial regression model with the load sequence as a single input. It can quickly capture the driving influence of demand on electricity prices. The model is constructed by selecting a polynomial form, where the order is adaptively selected based on cross-validation of historical data. The cross-validation operation principle is to divide the historical dataset into a training subset and a validation subset, and select the order with the minimum validation error after multiple fittings. The model training uses the least squares method to solve the regression coefficients. The least squares operation principle is to minimize the sum of the squared residuals of the polynomial term matrix and the historical electricity price vector, and calculate the coefficient vector through the matrix pseudo-inverse. The output reference curve operation principle is to substitute the current load sequence into the polynomial at each time point, multiply each term by the corresponding coefficient, and then sum them to ensure that the reference curve reflects the benchmark role of the load trend. Thus, the reference curve serves as a reliable anchor point for fusion correction, making it easier to expose the systemic bias of the first-round fusion curve.

[0080] 5-2. Based on the difference between the reference curve and the first-round fusion curve, a deviation vector is generated. This difference is a point-by-point subtraction operation that can quantify the spatiotemporal distribution of fusion error. First, at each time point, the reference curve value is subtracted from the first-round fusion curve value to generate a deviation sequence as a deviation vector. The operation principle is to realize the element-level calculation of the deviation vector through synchronous subtraction, ensuring that the deviation vector captures positive and negative offsets to reflect overfitting or underfitting, so that the deviation vector directly guides confidence adjustment to reduce fusion mismatch in a targeted manner.

[0081] 5-3. The confidence map is refreshed using the deviation vector. This refresh is a deviation-driven exponential decay method that can dynamically reduce the confidence score of large deviation regions. First, the local integral of the deviation vector is calculated. The local integral is obtained by accumulating the absolute value of the deviation vector point by point within a sliding window and multiplying it by the time interval. Then, the negative value of this integral is taken as the input of the exponential basis, and the reduction factor is calculated by substituting the negative integral into the natural exponent to ensure that the factor is less than one and inversely proportional to the deviation intensity. Next, the reduction factor is multiplied point by point by the corresponding element of the confidence score array to generate an updated array that covers the original confidence map. The calculation idea is to respond to the dynamic deviation through the decay mechanism to ensure that the confidence map adapts to the iteration requirements. After the refresh, the next step S4 is triggered immediately, and the fusion is re-executed using the updated confidence map to form a closed-loop correction to gradually converge the curve accuracy.

[0082] 5-4. Monitoring whether the deviation vector converges to a predetermined limit. This convergence judgment is a magnitude comparison operation that can evaluate the iteration termination condition. First, the overall magnitude of the deviation vector is calculated. The overall magnitude is the square root of the sum of the squares of all elements of the deviation vector, quantifying the total deviation amplitude. The predetermined limit is adaptively set by the median value of the integral of historical deviation magnitudes. The principle of the integral median value calculation is to sort the integrals of multiple historical magnitudes and take the middle value to ensure that the limit reflects the typical convergence level. If the overall magnitude is greater than the predetermined limit, the iteration continues. If it is less than or equal to the predetermined limit, the current fusion curve is output as the prediction curve, and the confidence evolution trajectory is written back. This trajectory is implemented by serializing and storing the confidence score array after each refresh, which facilitates subsequent retrieval and tracing of the correction path. Thus, the prediction curve integrates multiple rounds of optimization after convergence to achieve the final output of noise minimization and trend accuracy.

[0083] The forecast curve is the final output generated through a multimodal data fusion method during the forecasting of spot electricity prices in the new energy power market. This curve represents the predicted value of future spot electricity prices and integrates the contributions of multimodal data such as wind and solar power output sequences, load sequences, and matching signal sequences. Through processing steps S1 to S5, including instantaneous bandwidth division, confidence information extraction, dynamic revision, differential overlay, and deviation correction, the forecast curve is ensured to capture electricity price trends while suppressing high-frequency noise interference, thereby providing a reliable reference for market clearing and decision-making.

[0084] This invention separates steady-state and disturbance signals by dividing the instantaneous bandwidth of a multimodal sequence, avoiding direct noise interference in subsequent processes. Next, it extracts and dynamically revises confidence information to ensure accurate assessment of the reliability of different components. Then, in the fusion stage, it prioritizes the processing of high-confidence components and smooths low-confidence components to preserve the core trend while reducing fluctuation deviations. Finally, it iteratively corrects deviations using a load benchmark until the output prediction curve is converged. These steps collaborate: the front-end partitioning provides clean input, the mid-stage revision and fusion interactively optimize weights, and the back-end correction refines the results through feedback loops. This effectively suppresses high-frequency noise, captures dynamic market changes, and improves the accuracy and stability of the prediction curve in renewable energy price forecasting, providing more reliable support for power trading decisions.

[0085] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0086] It should be noted that the system of the present invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting a variety of hardware environments and usage requirements.

[0087] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0088] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely to distinguish one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0089] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A spot electricity price forecasting method based on multimodal data fusion, characterized in that, Including the following steps: S1: Real-time extraction of multi-modal sequences such as wind and solar power output, load and matching signals, adaptive division into steady-state bundles and disturbance bundles based on instantaneous bandwidth and then writing them into the component pool; S2: Extract confidence information from the short-window reconstruction error curve of each sequence in the component pool and write it into the confidence mapping table; S3: Based on phase consistency characteristics and amplitude mutation characteristics, a revision factor is output through a multi-core support vector machine to refresh the confidence mapping table level in real time; S4: First, overlay the high-confidence components according to the refreshed confidence mapping table, and then perform phase stretching smoothing on the low-confidence components before overlaying to generate the first round of fusion curve; S5: The reference curve is output by the load-driven fast reference model. The difference between the curve and the first-round fusion curve is used to obtain the deviation vector. The confidence mapping table is refreshed with the deviation vector and the next round of step S4 is triggered. When the deviation vector converges to the predetermined limit, the prediction curve is output and the confidence evolution trajectory is written back for subsequent retrieval.

2. The spot electricity price forecasting method based on multimodal data fusion according to claim 1, characterized in that, Step S1 includes the following: The power output sequence, load sequence, and matching signal sequence are updated every second through the data interface to form a continuous multimodal sequence set. The instantaneous bandwidth of each sequence in the multimodal sequence set is calculated using Hilbert transform to quantify the intensity of local fluctuations. An adaptive threshold is calculated based on the instantaneous bandwidth to determine the sequence segment assignment. Sequence segments with instantaneous bandwidth less than or equal to the threshold are assigned to the steady-state bundle, while those with bandwidth greater than the threshold are assigned to the disturbance bundle. The steady-state bundle and the disturbance bundle are written into the component pool.

3. The spot electricity price forecasting method based on multimodal data fusion according to claim 2, characterized in that, Step S2 includes the following: Each sequence from the steady-state and perturbation bundles is read from the component pool to form a set of sequences to be processed. The amplitude change rate of the local interval of the sequence is calculated to dynamically set the window length to adapt to the fluctuation scale. Hilbert transform is applied to the sequence sample points within the window to generate analytical signals and extract instantaneous amplitude and phase. The instantaneous amplitude is mapped according to the instantaneous frequency distribution to generate a local Hilbert spectrum. The phase integral and amplitude product of the selected principal mode component are summed through inverse transform to reconstruct the approximate sequence. The root mean square error is used to calculate the reconstruction error curve to quantify the deviation between sequences.

4. The spot electricity price forecasting method based on multimodal data fusion according to claim 3, characterized in that, Step S2 also includes the following: The confidence score is extracted by taking the reciprocal of the error curve through integral reconstruction, and the sequence window corresponding to the confidence score is written into the confidence mapping table.

5. The spot electricity price forecasting method based on multimodal data fusion according to claim 4, characterized in that, Step S3 includes the following: The original sequence corresponding to the confidence score array of each sequence is read from the confidence mapping table to form a feature extraction sequence set; the reverse sequence of the generated sequence is normalized by cross-correlation operation and the maximum value is taken to extract the phase overlap to quantify the phase matching degree; The ratio of the difference between the mean and the high quantile value of the amplitude sequence is calculated in the same window to extract the amplitude jump rate and quantify the amplitude deviation.

6. The spot electricity price forecasting method based on multimodal data fusion according to claim 5, characterized in that, Step S3 also includes the following: Using phase overlap and velocity jump rate as input vectors, a multi-kernel support vector machine is trained to minimize risk, solve for support vectors, and calculate the inner product, outputting the dynamics of the revision factor response features; the revision factor is multiplied by the confidence score array to refresh the confidence mapping table.

7. The spot electricity price forecasting method based on multimodal data fusion according to claim 6, characterized in that, Step S4 includes the following: Read the confidence score array of all sequences from the refreshed confidence mapping table to form a component classification set. Sequences with confidence scores higher than a preset threshold are classified as high-confidence components, and sequences with confidence scores lower than or equal to the preset threshold are classified as low-confidence components.

8. The spot electricity price forecasting method based on multimodal data fusion according to claim 7, characterized in that, Step S4 also includes the following: The high-confidence components are added point by point to generate a preliminary curve, thereby accumulating a reliable trend; The phase sequence of the low-confidence component is extracted, and the inverse of the local curvature integral is calculated to obtain the phase scaling factor multiplied by the phase cosine to reconstruct the smoothed sequence, so as to compress the oscillation amplitude. The smoothed low-confidence component is added to the preliminary curve to generate the first-round fusion curve to integrate the component contributions.

9. The spot electricity price forecasting method based on multimodal data fusion according to claim 8, characterized in that, Step S5 includes the following: Activate the load-driven fast reference model, capture the load trend with multinomial regression, output the reference curve, subtract it point by point from the first round of fusion curve to generate a deviation vector to quantify the error distribution; use the local integral negative exponential decay confidence score array of the deviation vector to refresh the confidence mapping table, and trigger the next round step S4 after refreshing, and re-execute the fusion using the updated confidence mapping table.

10. The spot electricity price forecasting method based on multimodal data fusion according to claim 9, characterized in that, Step S5 also includes the following: Calculate the overall magnitude of the deviation vector, compare it with the predetermined limit set by the median value of the historical magnitude integral, if the overall magnitude is greater than the predetermined limit, continue iterating, if it is less than or equal to the predetermined limit, output the current fusion curve as the prediction curve, and write back the confidence evolution trajectory.

Images